[codex] Add AF3 layout and Boltz-2 parser support

[DimaMolod]

Summary

This PR fixes the two parser compatibility gaps reported in issues #3 and #11:

• adds direct Boltz-2 prediction-directory support (<input>_model_<rank>.cif, confidence_<input>_model_<rank>.json, optional PAE/pLDDT NPZ files)
• teaches the AF3 parser to detect both AlphaPulldown/normalized AF3 output and current vanilla DeepMind AlphaFold3 output with <job>_ranking_scores.csv and prefixed per-sample files
• adds a real Boltz-2 fixture from the best readable 6OGE_ABC_DSSO_CDI seed/model plus focused regression tests
• bumps the package version to 1.0.2

Closes #3.
Fixes #11.

Notes

Vanilla AlphaFold3 writes one sanitized job directory under the selected --output_dir. Users passing the parent output directory should still run AlphaJudge with --recursive, or pass the job directory directly.

Boltz-2 protein-ligand outputs can include trailing small-molecule tokens in PAE/pLDDT arrays. AlphaJudge now trims larger Boltz-2 arrays to the protein/nucleic residue block it scores.

Validation

PYTHONPATH=src python -m py_compile \
  src/alphajudge/parsers/af3.py \
  src/alphajudge/parsers/boltz.py \
  src/alphajudge/parsers/__init__.py \
  test/test_parsers_and_runner.py

OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 NUMEXPR_NUM_THREADS=1 \
PYTHONPATH=src conda run -n alphajudge pytest \
  test/test_parsers_and_runner.py::test_af3_parser_accepts_official_prefixed_layout \
  test/test_parsers_and_runner.py::test_boltz2_parser_processes_ranked_prediction_dir \
  --tb=short -q

Result: 2 passed in 12.54s.

Additional parser smoke checks:

• AlphaPulldown-style AF3 fixture: af3 5 seed-19698302_sample-1
• Real Boltz-2 fixture: boltz2 6OGE_ABC_DSSO_CDI_Boltz2_model_0 (1058, 1058) 1058 0.8897088170051575

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 00959778bf

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Raise on unknown AF3 confidence schema

Falling back to an all-100 PAE matrix here silently turns missing/invalid AF3 confidence files into apparently valid results, which can produce misleading interface metrics instead of failing fast. This is triggered whenever confidences.json/summary_confidences.json is absent or in an unsupported format (for example, a layout mismatch where _read_json returns {}), and the run is then scored with fabricated PAE values rather than being skipped with an error as before.

Useful? React with 👍 / 👎.